deadrisk

Analysis

The Labs Now Charge the Same. The Bills Still Differ.

Between 2 September and 7 October OpenAI's and Anthropic's list prices met on every tier. Since then the contest has moved to how many tokens a task burns and who gives discounts, and on the record so far OpenAI is ahead on both.

ANTGOOGL
Coverage intelligence, not investment advice — methodology.

Three tiers, one price list

OpenAI and Anthropic now list the same prices at the top, the middle and the bottom. In July they matched on none. On 10 July OpenAI listed its top model, GPT-5.6 Sol, at $5 and $30 per million input and output tokens, Terra at $2.50 and $15 and the small Luna at $1 and $6. Anthropic’s top model, Fable 5, listed at $10 and $50. Its Sonnet 5 had launched on 30 June at $2 and $10, with a rise to $3 and $15 planned for September that it cancelled in August. Haiku 4.5 listed at $1 and $5.

OpenAI cut first. It took Luna to $0.20 and $1.20 and Terra down 20% on 30 July and Sol to $4 and $20 on 21 August. Then the tiers met. Anthropic kept Fable 5.1 at Fable 5’s $10 and $50 on 2 September and cut its cache-read price by 75%. Two days later OpenAI priced GPT-6 Astra, a new tier above Sol, at the same $10 and $50: at the top, OpenAI’s list price rose to meet Anthropic’s. On 22 September GPT-6 Sol came out at $2 and $10, Sonnet’s price since June, and GPT-6 Luna at $0.10 and $0.50. On 7 October Anthropic’s Haiku 5.5 matched Luna at $0.10 and $0.50 for requests up to 100,000 tokens, and $0.50 and $2.50 above that, and Anthropic halved Sonnet 5.5’s cache-read price.

35 days
From Fable 5.1's price on 2 September to Haiku 5.5's on 7 October: the span in which TEXXR's records show the two labs' list prices meeting on every tier
In July they matched on none: Fable 5 at $10/$50 against GPT-5.6 Sol at $5/$30. OpenAI moved to Anthropic's price at the top (up) and in the middle (down); Anthropic moved to OpenAI's at the bottom (down)
Tier OpenAI Anthropic
Top GPT-6 Astra, $10 / $50 Fable 5.1, $10 / $50
Middle GPT-6.1 Sol, $2 / $10 Sonnet 5.5, $2 / $10
Small GPT-6 Luna, $0.10 / $0.50 Haiku 5.5, $0.10 / $0.50 (to 100K tokens)

OpenAI kept adding cheaper lines: a Decisions API at $0.10 per million input tokens with no output charge, by TEXXR’s record of the announcement, and on 26 October GPT-6.1 Terra at $1 and $5. It also halved the API credits it gives per dollar in late September.

Same price, different bill

A list price says little without the tokens a task burns. As The Decoder reports Artificial Analysis’s figures, Astra used about 27,000 output tokens per task, Fable 5.1 about 78,000 and Opus 5.5, at maximum effort, about 119,000. At the same list price, Fable 5.1’s output for a task would cost about 2.9 times Astra’s, derived. Anthropic priced Opus 5.5 at $4 and $20 and said it costs about 40% less to run than Opus 5; Artificial Analysis judged its cost per task in line with Opus 5’s, not cheaper. It put Sonnet 5.5 at maximum effort at $7.60 a task, about 50% above Sonnet 5. Google’s Gemini 4 Argon lists at $2 and $10 as an introductory price, with $4 and $20 to follow on a date it has not given.

Who gives discounts

The discounts ran OpenAI’s way. The Information’s sources said Anthropic ended discounts, which typically reach about 15%, once customers pass their usage limits. Replit’s chief executive said on 16 October that it no longer works with Anthropic, citing cost, partnership and performance, and that “ChatGPT is better now”; a source told TechCrunch the partnership continues. The Information’s source said Replit moved most of its spend to OpenAI for more than 30% off, in a deal worth more than $100 million a year.

Spending followed. OpenRouter said on 16 September that OpenAI had out-earned Anthropic on its platform in a week for the first time since February 2024. The WSJ reported OpenRouter data showing business spending on the two about even in September, against a share above 75% for Anthropic in January.

Max Woolf’s timeline of the cuts reads them as a response to open-weight rivals and low switching costs, and notes that both labs are close to listings, “which raises further questions about the price cuts.” Artificial Analysis ranked Moonshot’s Kimi K3.1 Pro fourth on its index; Moonshot says the weights will be public by 4 November. Not every seller is cutting: DeepSeek raised V4 prices from 16 August, taking V4-Flash output to $0.66 off peak and $1.32 at peak, from $0.28.

What it does to the targets

Anthropic told investors its third-quarter revenue was $16.8 billion, the FT’s sources said; the Anthropic dossier derives about $16.9 billion from the S-1. Reuters’ sources put its 2028 projection at $190 billion to $200 billion, about three times that quarter annualised, derived. A leaked OpenAI presentation had revenue growing from $36 billion this year to $350 billion in 2030. Its annualised revenue is reported at about $70 billion by Axios and at nearly $50 billion in what it told investors, per the FT; against the lower figure the 2030 plan is about seven times, derived.

OpenAI cut its own valuation target to $1.25 trillion from $1.4 trillion after investor resistance, the FT’s sources said. Some Anthropic investors lowered theirs, discussing about $1.5 trillion rather than $2 trillion after slower growth, Reuters reported. No report ties either change to prices.

Revenue is price times tokens. Where a price fell by two thirds, holding revenue flat takes three times the tokens; a target three times today’s revenue would take about nine times. These figures are derived, at a constant mix, and ignore discounts and moves between labs. Apollo’s daily note says that outside tech there is no financial payoff from AI yet, and that mainstream firms treat tokens as a cost that weighs on margins. It says nothing about the labs’ prices.

What the forecast does with it

DeadRisk’s forecast has a “Commodity intelligence” preset. Against the default world it changes four lab inputs. Frontier work falls to 10% of AI work by 2032, not 30%. Its price is capped at 1.5 times its compute cost, not 3 times; the page equates 3 times with a 67% gross margin, so 1.5 times is 33%. By 2032 its price reaches 15% of a person’s cost, not 30%. And by 2030 compute rents recover half the fleet’s full cost when supply and demand balance, not all of it. The model starts 2026 with half of AI work at frontier prices and buyer spending near $170 billion, its own estimate. This piece quotes no output; load the preset to see who ends ahead of their cost of capital. The token ceiling trend sets out the question. On list prices, October looks like that world below the top tier; on cost per task, the labs are not yet equal.

What to watch

  • 4 November. Moonshot’s date for Kimi K3.1 Pro’s weights. Woolf counted 29 hosts on OpenRouter for DeepSeek’s latest Flash, against 13 for the earlier V4 Flash in May.
  • Tokens per task. Whether Anthropic’s next models close the gap with Astra that Artificial Analysis measured.
  • Week of 16 November. The WSJ’s sources put Anthropic’s listing there, per its dossier; an amended filing should state third-quarter revenue.
  • Argon’s second price. Google gave no date for the rise to $4 and $20; see the Google dossier.
  • The OpenAI round. Whether it closes at $1.25 trillion.

Sources

Dates are the publisher’s where the page shows one, otherwise TEXXR’s.

Disclosure: this piece was drafted with Claude, a model made by Anthropic, one of the companies it covers.